Catálogo de publicaciones - libros

While features used for texture analysis have been successfully used in some biometric applications, only quite few works have considered them in facial image analysis. Texture-based region descriptors can be very useful in recognizing faces and facial expressions, detecting faces and different facial components, and in other face related tasks. This paper demonstrates this issue by considering the local binary pattern (LBP) as an example of texture-based approach and showing its efficiency in facial image analysis.

Active Shape Model (ASM) has been recognized as one of the typical methods for image understanding. Simply speaking, it iterates two steps: profile-based landmarks local searching, and statistics-based global shape modeling. We argue that the simple 1D profile matching may not localize landmarks accurately enough, and the unreliable localized landmarks will mislead the following shape matching. Considering these two problems, we propose to enhance ASM from two aspects: (1) in the landmarks local searching step, we introduce more efficient AdaBoost method to localize some salient landmarks instead of the relatively simple profile matching as in the traditional ASMs; (2) in the global shape modeling step, the confidences of the landmark localization are exploited to constrain the shape modeling and reconstruction procedure by not using those unreliably located landmarks to eliminate their negative effects. We experimentally show that the proposed strategies can impressively improve the accuracy of the traditional ASMs.

This paper proposes a new method for personal identity verification based the analysis of face images applying One Class Support Vector Machines. This is a recently introduced kernel method to build a unary classifier to be trained by using only positive examples, avoiding the sensible choice of the impostor set typical of standard binary Support Vector Machines. The features of this classifier and the application to face-based identity verification are described and an implementation presented. Several experiments have been performed on both standard and proprietary databases. The tests performed, also in comparison with a standard classifier built on Support Vector Machines, clearly show the potential of the proposed approach.

The problem of illumination makes the face recognition still an unsolved problem. A new normalization method was presented which is illumination invariant in face recognition. The histogram equalization was applied to improve the performances of the AT (Affine Transform) algorithm. A novel illumination normalization method was introduced. Using the information of the distribution of the histogram, the method makes the combination of the AT algorithm and the ICR (Illumination Compensation based on Multiple Regression Model) algorithm smoothly. Experiments reveals that our proposed algorithm is illumination invariant and achieves better preprocess result while the recognition rate has been evidently improved.

This paper investigates the score normalization technique for enhancing the performance of face authentication. We firstly discuss the thresholding approach for face authentication and put forward the “score variation” problem. Then, two possible solutions, Subject Specific Threshold (SST) and Score Normalization (SN), are discussed. But SST is obviously impractical to many face authentication applications in which only a single example face image is available for each subject. Fortunately, we have theoretically shown that, in such cases, score normalization technique may approximately approach the SST by using a uniform threshold. Experiments on both the FERET and CAS-PEAL face database have shown the effectiveness of SN for different face authentication methods including Correlation, Eigenface, and Fisherface.

Though AdaBoost has been widely used for feature selection and classifier learning, many of the selected features, or weak classifiers, are redundant. By incorporating mutual information into AdaBoost, we propose an improved boosting algorithm in this paper. The proposed method fully examines the redundancy between candidate classifiers and selected classifiers. The classifiers thus selected are both accurate and non-redundant. Experimental results show that the strong classifier learned using the proposed algorithm achieves a lower training error rate than AdaBoost. The proposed algorithm has also been applied to select discriminative Gabor features for face recognition. Even with the simple correlation distance measure and 1-NN classifier, the selected Gabor features achieve quite high recognition accuracy on the FERET database, where both expression and illumination variance exists. When only 140 features are used, the selected features achieve as high as 95.5% accuracy, which is about 2.5% higher than that of features selected by AdaBoost.

In this paper, we propose an automatic method to set up a 3D morphable face model. The main idea is to combine the algorithm of Fast-AAM with Thin Plate Splines (TPS) for 3D data alignment. Our method is better than the algorithm of optical flow proposed in [3][5] because it avoids the local minima problem, and also it is better than the adaptive multi-resolution fitting algorithm [8][9] in 3D alignment because it is fully automatic and the latter needs to be manually initialized. Result shows that our method is practicable in speed and accuracy.

We propose two new classifiers, one based on the classical Procrustes analysis and the other on the Moore-Penrose inverse in the context of face recognition. The Procrustes based classifier has recognition rates of 97.5%, 96.19%, 71.40% and 96.22% for the ORL, YALE, GIT and the FERET database respectively. The Moore-Penrose classifier has comparative recognition rates of 98%, 99.04%, 87.40% and 96.22% for the same databases. In addition to these classifiers, we also propose new parameters that are useful for comparing classifiers based on their discriminatory power and not just on their recognition rates. We also compare the performance of our classifiers with the baseline PCA and LDA techniques as well as the recently proposed discriminative common vectors technique for the above face databases.

Though authentication using biometric techniques has conveniences for people, security problems like the leakage of personal bio-information would be serious. Even if cancelable biometric is a good solution for the problems, only a few biometric authentication scheme with cancelable feature has been published. In this paper, we suggest a face authentication scheme with two security factors: password and face image. Using matching algorithm in the permuted domain, our scheme is designed to be cancelable in the sense that templates that is composed of permutation and weight vector can be changed freely.

A set of local feature descriptors for fingerprints is proposed. Minutia points are detected in a novel way by complex filtering of the structure tensor, not only revealing their position but also their direction. Parabolic and linear symmetry descriptions are used to model and extract local features including ridge orientation and reliability, which can be reused in several stages of fingerprint processing. Experimental results on the proposed technique are presented.